saen_dev Profile Banner
saen Profile
saen

@saen_dev

Followers
59
Following
174
Media
32
Statuses
437

AI Engineer building production agents with LangGraph & DSPy | Sharing real implementations & technical insights

UAE , Dubai
Joined October 2023
Don't wanna be here? Send us removal request.
@saen_dev
saen
10 days
AI will not replace you. A person using AI will replace you....
0
0
2
@saen_dev
saen
2 hours
The pattern is clear: Text → ChatGPT dominated Images → DALL-E changed the game Code → Copilot revolutionized workflows Now music is next. When OpenAI enters a space, they redefine the market. Suno and Udio just got their "ChatGPT moment" 🎵
@VraserX
VraserX e/acc
10 hours
OpenAI just dropped the bass. The company that changed how we write, code, and create images is now tuning its sights on music. OpenAI is reportedly building an AI music generator that could rival Suno and Udio. Think about that for a second. ChatGPT turned words into meaning.
0
0
0
@saen_dev
saen
4 hours
@unwind_ai_ WAIT ByteDance really just dropped a 0.3B parameter OCR model that reads like humans?? 🤯 the fact that it analyzes layout FIRST then parses in parallel is actually genius level architecture. and it's fully open source??? this is about to change document processing forever 📄✨
0
1
0
@saen_dev
saen
5 hours
@Web3Primee okayyy this stack is actually clean af 🔥 Next.js 15 + TypeScript is *chef's kiss* and the Solidity integration with OpenZeppelin?? love to see proper security practices the localStorage with auto-restore is such a smart UX move. production-ready fr 💻✨
0
1
0
@saen_dev
saen
7 hours
3/ What's next in 2025? → AI AGENTS working autonomously (not just chatbots) → Inference costs: $20 → $0.07 per million tokens → Europe's AI Act enforced from August → Scientific breakthroughs accelerating We're witnessing history unfold in real-time 🌍
0
0
0
@saen_dev
saen
7 hours
2/ ChatGPT Atlas launched - OpenAI's AI browser with GPT-5 integrated + Claude gains MEMORY (remembers your conversations) + Google's DeepSomatic detects cancer variants with unprecedented accuracy + Anthropic's Claude Sonnet 4.5 crushes coding benchmarks The race is ON 🔥
1
0
0
@saen_dev
saen
7 hours
🚀 AI just had its BIGGEST week in October 2025 Here's what's reshaping the entire industry (thread 🧵) 1/ OpenAI + AMD: Multi-BILLION $ deal for 6 gigawatts of AI chips - challenging Nvidia's dominance First deployment: AMD MI450 GPUs in 2026 This is a game-changer 💻
1
0
0
@saen_dev
saen
18 hours
Building production agents is different than demos. Key lessons: ✅ Parallel > Sequential ✅ Checkpoint everything ✅ Right model for right task ✅ Stream responses ✅ Eval from day 1 ✅ Optimize prompts systematically What mistakes have you made? Reply below 👇
0
0
0
@saen_dev
saen
18 hours
Bonus: Not using DSPy for prompt optimization Hand-crafted prompts = guesswork DSPy's BootstrapFewShot = systematic improvement Took our accuracy from 79% to 91% in 3 hours of training. The math doesn't lie. Let the framework do the work.
1
0
0
@saen_dev
saen
18 hours
5/ No evaluation framework from day one We didn't track agent performance systematically. Bad outputs went unnoticed for weeks. LangSmith + custom evals changed everything: - Caught regressions immediately - A/B tested prompt changes - Reduced hallucinations by 68%
1
0
0
@saen_dev
saen
18 hours
4/ Ignoring streaming Batch responses felt broken to users. Average perceived wait time: 8 seconds Abandonment rate: 23% Added streaming: Perceived wait: 2 seconds Abandonment: 7% Users don't mind waiting if they see progress.
1
0
0
@saen_dev
saen
18 hours
3/ Over-relying on expensive models Started with GPT-4 for everything. Monthly cost: $8,200 Switched to Haiku 4.5 for routing + GPT-4 only for final decisions. New cost: $2,100 Same quality. 75% cost reduction. Just needed proper prompt engineering.
1
0
0
@saen_dev
saen
18 hours
2/ Not implementing proper checkpointing Lost hours of agent work when servers restarted. One customer's complex workflow died at 95% completion. Solution: Persistent state with checkpointers. Now agents resume exactly where they left off. Game changer.
1
0
0
@saen_dev
saen
18 hours
1/ Running agents sequentially instead of parallel We had 4 agents checking different data sources. Sequential: 12s average Parallel with proper state management: 2.8s Lesson: LangGraph's graph design isn't just fancy — it's fundamentally faster in production.
1
0
0
@saen_dev
saen
18 hours
5 LangGraph mistakes that cost us $15K in our first production month. After 6 months running agents at scale, here's what we learned the hard way 🧵
1
0
0
@saen_dev
saen
18 hours
This is why LangGraph's graph design shines. Parallel agent execution eliminates the sequential bottleneck. With proper state management + routing, run multiple agents concurrently without the wait. Same reasoning depth, fraction of latency. Already in production. 🧠
@vishalxai
Vishal Singh | AI/ML Engineer
2 months
Traditionally, we scaled AI with: •More data •Bigger training compute •Test-time reasoning (slower, deeper) Now → parallel agents let us scale performance without making users wait.
0
0
1
@saen_dev
saen
19 hours
OpenLiberty 25.0.0.11 shipping MCP Server support is huge. Model Context Protocol is becoming the standard bridge between LLMs and real systems. Ultra-fast I/O + MCP = agents that can actually interact with enterprise infrastructure at scale 🔥
@OpenLibertyIO
openlibertyio
22 hours
🚀 #OpenLiberty 25.0.0.11-beta is here—bridging #AI & ultra-fast I/O! This release introduces two powerful features: ✅ Model Context Protocol (#MCP) Server ✅ Netty-based #HTTP transport 🔗 Dive into the details: https://t.co/59YDsH2oiX #LLM
0
0
0
@saen_dev
saen
19 hours
AI + Blockchain is the convergence everyone's talking about. Decentralized ML models, agents operating on-chain, AI-governed smart contracts - this is where autonomous systems get real economic power. We're just starting to see what becomes possible at the intersection 🔥
0
0
0
@saen_dev
saen
19 hours
Karpathy's right. Today's agents have hard limits - can't operate systems, blend modalities, or learn continuously. This is why we focus on reliability over hype. Intern-level AI requires infrastructure + iteration, not breakthroughs. The march of nines is the real game 🔥
@Mahesh_Lambe
Mahesh Lambe
3 days
This is wild. Karpathy on agents: today’s coding agents can’t operate computers, integrate multiple modalities or learn continuously. Building reliable intern-level AI will take years of infrastructure and the 'march of nines' improvements, not a sudden breakthrough.
0
0
0
@saen_dev
saen
19 hours
Google's $10B+ chip deal with Anthropic is massive. This is the infrastructure race heating up. When companies are betting billions on chip partnerships, it means AI at scale is real. The winners will be those building agents that can leverage this compute efficiently 🔥
@FintechTvGlobal
FINTECH.TV
20 hours
Here’s today’s Market Movers: AI Daily — your 2-minute rundown on the biggest AI headlines. Google’s $10B+ chip deal with Anthropic, OpenAI’s latest acquisition, Intel’s post-earnings rally, and why Nvidia’s investing in battery recycling firm Redwood Materials.
0
0
0
@saen_dev
saen
19 hours
Agentic browsers are the real evolution. We've been building towards this - agents that don't just observe but actually execute tasks. Book flights, compare prices, handle workflows. This is where agents move from fun experiments to replacing manual work
@AskPerplexity
Ask Perplexity
19 hours
@ASIsoftware @AravSrinivas @perplexity_ai Agentic browsers are the next evolution of web browsing — they don’t just show you websites, they actually *act* on your behalf. Instead of clicking and typing, you can tell them to “book a flight” or “compare hotel prices,” and the browser’s built‑in AI agents handle the
0
0
0